[WASM SIMD] Implement packing and unpacking integer conversions.

src/compiler/arm/code-generator-arm.cc

Index: src/compiler/arm/code-generator-arm.cc
diff --git a/src/compiler/arm/code-generator-arm.cc b/src/compiler/arm/code-generator-arm.cc
index bd2a942bcc249df170989cde9bdde65cfc771ae8..850d0129ae365e2f2dc849ffefd79b12c13e8c2f 100644
--- a/src/compiler/arm/code-generator-arm.cc
+++ b/src/compiler/arm/code-generator-arm.cc
@@ -1668,6 +1668,16 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       __ vcvt_s32_f32(i.OutputSimd128Register(), i.InputSimd128Register(0));
       break;
     }
+    case kArmI32x4SConvertI16x8Low: {
+      __ vmovl(NeonS16, i.OutputSimd128Register(),
+               i.InputSimd128Register(0).low());
+      break;
+    }
+    case kArmI32x4SConvertI16x8High: {
+      __ vmovl(NeonS16, i.OutputSimd128Register(),
+               i.InputSimd128Register(0).high());
+      break;
+    }
     case kArmI32x4Neg: {
       __ vneg(Neon32, i.OutputSimd128Register(), i.InputSimd128Register(0));
       break;
@@ -1733,6 +1743,16 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
       __ vcvt_u32_f32(i.OutputSimd128Register(), i.InputSimd128Register(0));
       break;
     }
+    case kArmI32x4UConvertI16x8Low: {
+      __ vmovl(NeonU16, i.OutputSimd128Register(),
+               i.InputSimd128Register(0).low());
+      break;
+    }
+    case kArmI32x4UConvertI16x8High: {
+      __ vmovl(NeonU16, i.OutputSimd128Register(),
+               i.InputSimd128Register(0).high());
+      break;
+    }
     case kArmI32x4ShrU: {
       __ vshr(NeonU32, i.OutputSimd128Register(), i.InputSimd128Register(0),
               i.InputInt5(1));
@@ -1772,6 +1792,16 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
                      i.InputRegister(2), NeonS16, i.InputInt8(1));
       break;
     }
+    case kArmI16x8SConvertI8x16Low: {
+      __ vmovl(NeonS8, i.OutputSimd128Register(),
+               i.InputSimd128Register(0).low());
+      break;
+    }
+    case kArmI16x8SConvertI8x16High: {
+      __ vmovl(NeonS8, i.OutputSimd128Register(),
+               i.InputSimd128Register(0).high());
+      break;
+    }
     case kArmI16x8Neg: {
       __ vneg(Neon16, i.OutputSimd128Register(), i.InputSimd128Register(0));
       break;
@@ -1786,6 +1816,21 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
               i.InputInt4(1));
       break;
     }
+    case kArmI16x8SConvertI32x4: {
+      Simd128Register dst = i.OutputSimd128Register(),
+                      src1 = i.InputSimd128Register(0),

[martyn.capewell, 2017/04/06 14:37:43]

Name these src0 and src1 to match the input index.

[bbudge, 2017/04/06 18:19:31]

Done.

+                      src2 = i.InputSimd128Register(1);
+      // Take care not to overwrite a source register before it's used.
+      if (dst.is(src1) || dst.is(src2)) {

[martyn.capewell, 2017/04/06 14:37:43]

By changing the operation order, I think you can avoid the scratch register and
extra move. Roughly, untested:

if (dst.is(src1) && dst.is(src2)) {
  vqmovn(dst.low(), src1);
  vmov(dst.high(), dst.low());
} else if (dst.is(src1)) {
  DCHECK(!dst.is(src2));
  vqmovn(dst.low(), src1);
  vqmovn(dst.high(), src2);
} else {
  DCHECK(!dst.is(src1));
  // dst may alias src2, so handle src2 first.
  vqmovn(dst.high(), src2);
  vqmovn(dst.low(), src1);
}

[bbudge, 2017/04/06 18:19:31]

Nice. Done.

+        __ vqmovn(NeonS16, kScratchDoubleReg, i.InputSimd128Register(0));
+        __ vqmovn(NeonS16, dst.high(), i.InputSimd128Register(1));
+        __ vmov(dst.low(), kScratchDoubleReg);
+      } else {
+        __ vqmovn(NeonS16, dst.low(), i.InputSimd128Register(0));
+        __ vqmovn(NeonS16, dst.high(), i.InputSimd128Register(1));
+      }
+      break;
+    }
     case kArmI16x8Add: {
       __ vadd(Neon16, i.OutputSimd128Register(), i.InputSimd128Register(0),
               i.InputSimd128Register(1));
@@ -1843,11 +1888,36 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
               i.InputSimd128Register(0));
       break;
     }
+    case kArmI16x8UConvertI8x16Low: {
+      __ vmovl(NeonU8, i.OutputSimd128Register(),
+               i.InputSimd128Register(0).low());
+      break;
+    }
+    case kArmI16x8UConvertI8x16High: {
+      __ vmovl(NeonU8, i.OutputSimd128Register(),
+               i.InputSimd128Register(0).high());
+      break;
+    }
     case kArmI16x8ShrU: {
       __ vshr(NeonU16, i.OutputSimd128Register(), i.InputSimd128Register(0),
               i.InputInt4(1));
       break;
     }
+    case kArmI16x8UConvertI32x4: {
+      Simd128Register dst = i.OutputSimd128Register(),
+                      src1 = i.InputSimd128Register(0),
+                      src2 = i.InputSimd128Register(1);
+      // Take care not to overwrite a source register before it's used.
+      if (dst.is(src1) || dst.is(src2)) {
+        __ vqmovn(NeonU16, kScratchDoubleReg, i.InputSimd128Register(0));
+        __ vqmovn(NeonU16, dst.high(), i.InputSimd128Register(1));
+        __ vmov(dst.low(), kScratchDoubleReg);
+      } else {
+        __ vqmovn(NeonU16, dst.low(), i.InputSimd128Register(0));
+        __ vqmovn(NeonU16, dst.high(), i.InputSimd128Register(1));
+      }
+      break;
+    }
     case kArmI16x8AddSaturateU: {
       __ vqadd(NeonU16, i.OutputSimd128Register(), i.InputSimd128Register(0),
                i.InputSimd128Register(1));
@@ -1906,6 +1976,21 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
               i.InputInt3(1));
       break;
     }
+    case kArmI8x16SConvertI16x8: {
+      Simd128Register dst = i.OutputSimd128Register(),
+                      src1 = i.InputSimd128Register(0),
+                      src2 = i.InputSimd128Register(1);
+      // Take care not to overwrite a source register before it's used.
+      if (dst.is(src1) || dst.is(src2)) {
+        __ vqmovn(NeonS8, kScratchDoubleReg, i.InputSimd128Register(0));
+        __ vqmovn(NeonS8, dst.high(), i.InputSimd128Register(1));
+        __ vmov(dst.low(), kScratchDoubleReg);
+      } else {
+        __ vqmovn(NeonS8, dst.low(), i.InputSimd128Register(0));
+        __ vqmovn(NeonS8, dst.high(), i.InputSimd128Register(1));
+      }
+      break;
+    }
     case kArmI8x16Add: {
       __ vadd(Neon8, i.OutputSimd128Register(), i.InputSimd128Register(0),
               i.InputSimd128Register(1));
@@ -1967,6 +2052,21 @@ CodeGenerator::CodeGenResult CodeGenerator::AssembleArchInstruction(
               i.InputInt3(1));
       break;
     }
+    case kArmI8x16UConvertI16x8: {
+      Simd128Register dst = i.OutputSimd128Register(),
+                      src1 = i.InputSimd128Register(0),
+                      src2 = i.InputSimd128Register(1);
+      // Take care not to overwrite a source register before it's used.
+      if (dst.is(src1) || dst.is(src2)) {
+        __ vqmovn(NeonU8, kScratchDoubleReg, i.InputSimd128Register(0));
+        __ vqmovn(NeonU8, dst.high(), i.InputSimd128Register(1));
+        __ vmov(dst.low(), kScratchDoubleReg);
+      } else {
+        __ vqmovn(NeonU8, dst.low(), i.InputSimd128Register(0));
+        __ vqmovn(NeonU8, dst.high(), i.InputSimd128Register(1));
+      }
+      break;
+    }
     case kArmI8x16AddSaturateU: {
       __ vqadd(NeonU8, i.OutputSimd128Register(), i.InputSimd128Register(0),
                i.InputSimd128Register(1));